projects / libdai.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Even more cleanup of BBP code
[libdai.git] / src / bbp.cpp
1 /* Copyright (C) 2009 Frederik Eaton [frederik at ofb dot net]
2
3 This file is part of libDAI.
4
5 libDAI is free software; you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation; either version 2 of the License, or
8 (at your option) any later version.
9
10 libDAI is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
14
15 You should have received a copy of the GNU General Public License
16 along with libDAI; if not, write to the Free Software
17 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 */
19
20
21 #include <dai/bp.h>
22 #include <dai/bbp.h>
23 #include <dai/gibbs.h>
24 #include <dai/util.h>
25 #include <dai/bipgraph.h>
26
27
28 namespace dai {
29
30
31 using namespace std;
32
33
34 typedef BipartiteGraph::Neighbor Neighbor;
35
36
37 Prob unnormAdjoint( const Prob &w, Real Z_w, const Prob &adj_w ) {
38 assert( w.size() == adj_w.size() );
39 Prob adj_w_unnorm( w.size(), 0.0 );
40 Real s = 0.0;
41 for( size_t i = 0; i < w.size(); i++ )
42 s += w[i] * adj_w[i];
43 for( size_t i = 0; i < w.size(); i++ )
44 adj_w_unnorm[i] = (adj_w[i] - s) / Z_w;
45 return adj_w_unnorm;
46 // THIS WOULD BE ABOUT 50% SLOWER: return (adj_w - (w * adj_w).sum()) / Z_w;
47 }
48
49
50 std::vector<size_t> getGibbsState( const InfAlg &ia, size_t iters ) {
51 PropertySet gibbsProps;
52 gibbsProps.Set("iters", iters);
53 gibbsProps.Set("verbose", size_t(0));
54 Gibbs gibbs( ia.fg(), gibbsProps );
55 gibbs.run();
56 return gibbs.state();
57 }
58
59
60 size_t getFactorEntryForState( const FactorGraph &fg, size_t I, const vector<size_t> &state ) {
61 size_t f_entry = 0;
62 for( int _j = fg.nbF(I).size() - 1; _j >= 0; _j-- ) {
63 // note that iterating over nbF(I) yields the same ordering
64 // of variables as iterating over factor(I).vars()
65 size_t j = fg.nbF(I)[_j];
66 f_entry *= fg.var(j).states();
67 f_entry += state[j];
68 }
69 return f_entry;
70 }
71
72
73 #define LOOP_ij(body) { \
74 size_t i_states = _fg->var(i).states(); \
75 size_t j_states = _fg->var(j).states(); \
76 if(_fg->var(i) > _fg->var(j)) { \
77 size_t xij=0; \
78 for(size_t xi=0; xi<i_states; xi++) { \
79 for(size_t xj=0; xj<j_states; xj++) { \
80 body; \
81 xij++; \
82 } \
83 } \
84 } else { \
85 size_t xij=0; \
86 for(size_t xj=0; xj<j_states; xj++) { \
87 for(size_t xi=0; xi<i_states; xi++) { \
88 body; \
89 xij++; \
90 } \
91 } \
92 } \
93 }
94
95
96 void BBP::RegenerateInds() {
97 // initialise _indices
98 // typedef std::vector<size_t> _ind_t;
99 // std::vector<std::vector<_ind_t> > _indices;
100 _indices.resize( _fg->nrVars() );
101 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
102 _indices[i].clear();
103 _indices[i].reserve( _fg->nbV(i).size() );
104 foreach( const Neighbor &I, _fg->nbV(i) ) {
105 _ind_t index;
106 index.reserve( _fg->factor(I).states() );
107 for( IndexFor k(_fg->var(i), _fg->factor(I).vars()); k >= 0; ++k )
108 index.push_back( k );
109 _indices[i].push_back( index );
110 }
111 }
112 }
113
114
115 void BBP::RegenerateT() {
116 // _T[i][_I]
117 _T.resize( _fg->nrVars() );
118 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
119 _T[i].resize( _fg->nbV(i).size() );
120 foreach( const Neighbor &I, _fg->nbV(i) ) {
121 Prob prod( _fg->var(i).states(), 1.0 );
122 foreach( const Neighbor &J, _fg->nbV(i) )
123 if( J.node != I.node )
124 prod *= _bp_dual.msgM( i, J.iter );
125 _T[i][I.iter] = prod;
126 }
127 }
128 }
129
130
131 void BBP::RegenerateU() {
132 // _U[I][_i]
133 _U.resize( _fg->nrFactors() );
134 for( size_t I = 0; I < _fg->nrFactors(); I++ ) {
135 _U[I].resize( _fg->nbF(I).size() );
136 foreach( const Neighbor &i, _fg->nbF(I) ) {
137 Prob prod( _fg->factor(I).states(), 1.0 );
138 foreach( const Neighbor &j, _fg->nbF(I) )
139 if( i.node != j.node ) {
140 Prob n_jI( _bp_dual.msgN( j, j.dual ) );
141 const _ind_t &ind = _index( j, j.dual );
142 // multiply prod by n_jI
143 for( size_t x_I = 0; x_I < prod.size(); x_I++ )
144 prod[x_I] *= n_jI[ind[x_I]];
145 }
146 _U[I][i.iter] = prod;
147 }
148 }
149 }
150
151
152 void BBP::RegenerateS() {
153 // _S[i][_I][_j]
154 _S.resize( _fg->nrVars() );
155 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
156 _S[i].resize( _fg->nbV(i).size() );
157 foreach( const Neighbor &I, _fg->nbV(i) ) {
158 _S[i][I.iter].resize( _fg->nbF(I).size() );
159 foreach( const Neighbor &j, _fg->nbF(I) )
160 if( i != j ) {
161 Factor prod( _fg->factor(I) );
162 foreach( const Neighbor &k, _fg->nbF(I) ) {
163 if( k != i && k.node != j.node ) {
164 const _ind_t &ind = _index( k, k.dual );
165 Prob p( _bp_dual.msgN( k, k.dual ) );
166 for( size_t x_I = 0; x_I < prod.states(); x_I++ )
167 prod.p()[x_I] *= p[ind[x_I]];
168 }
169 }
170 // "Marginalize" onto i|j (unnormalized)
171 Prob marg;
172 marg = prod.marginal( VarSet(_fg->var(i), _fg->var(j)), false ).p();
173 _S[i][I.iter][j.iter] = marg;
174 }
175 }
176 }
177 }
178
179
180 void BBP::RegenerateR() {
181 // _R[I][_i][_J]
182 _R.resize( _fg->nrFactors() );
183 for( size_t I = 0; I < _fg->nrFactors(); I++ ) {
184 _R[I].resize( _fg->nbF(I).size() );
185 foreach( const Neighbor &i, _fg->nbF(I) ) {
186 _R[I][i.iter].resize( _fg->nbV(i).size() );
187 foreach( const Neighbor &J, _fg->nbV(i) ) {
188 if( I != J ) {
189 Prob prod( _fg->var(i).states(), 1.0 );
190 foreach( const Neighbor &K, _fg->nbV(i) )
191 if( K.node != I && K.node != J.node )
192 prod *= _bp_dual.msgM( i, K.iter );
193 _R[I][i.iter][J.iter] = prod;
194 }
195 }
196 }
197 }
198 }
199
200
201 void BBP::RegenerateInputs() {
202 _adj_b_V_unnorm.clear();
203 _adj_b_V_unnorm.reserve( _fg->nrVars() );
204 for( size_t i = 0; i < _fg->nrVars(); i++ )
205 _adj_b_V_unnorm.push_back( unnormAdjoint( _bp_dual.beliefV(i).p(), _bp_dual.beliefVZ(i), _adj_b_V[i] ) );
206 _adj_b_F_unnorm.clear();
207 _adj_b_F_unnorm.reserve( _fg->nrFactors() );
208 for( size_t I = 0; I < _fg->nrFactors(); I++ )
209 _adj_b_F_unnorm.push_back( unnormAdjoint( _bp_dual.beliefF(I).p(), _bp_dual.beliefFZ(I), _adj_b_F[I] ) );
210 }
211
212
213 void BBP::RegeneratePsiAdjoints() {
214 _adj_psi_V.clear();
215 _adj_psi_V.reserve( _fg->nrVars() );
216 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
217 Prob p( _adj_b_V_unnorm[i] );
218 assert( p.size() == _fg->var(i).states() );
219 foreach( const Neighbor &I, _fg->nbV(i) )
220 p *= _bp_dual.msgM( i, I.iter );
221 p += _init_adj_psi_V[i];
222 _adj_psi_V.push_back( p );
223 }
224 _adj_psi_F.clear();
225 _adj_psi_F.reserve( _fg->nrFactors() );
226 for( size_t I = 0; I < _fg->nrFactors(); I++ ) {
227 Prob p( _adj_b_F_unnorm[I] );
228 assert( p.size() == _fg->factor(I).states() );
229 foreach( const Neighbor &i, _fg->nbF(I) ) {
230 Prob n_iI( _bp_dual.msgN( i, i.dual ) );
231 const _ind_t& ind = _index( i, i.dual );
232 // multiply prod with n_jI
233 for( size_t x_I = 0; x_I < p.size(); x_I++ )
234 p[x_I] *= n_iI[ind[x_I]];
235 }
236 p += _init_adj_psi_F[I];
237 _adj_psi_F.push_back( p );
238 }
239 }
240
241
242 void BBP::RegenerateParMessageAdjoints() {
243 size_t nv = _fg->nrVars();
244 _adj_n.resize( nv );
245 _adj_m.resize( nv );
246 _adj_n_unnorm.resize( nv );
247 _adj_m_unnorm.resize( nv );
248 _new_adj_n.resize( nv );
249 _new_adj_m.resize( nv );
250 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
251 size_t n_i = _fg->nbV(i).size();
252 _adj_n[i].resize( n_i );
253 _adj_m[i].resize( n_i );
254 _adj_n_unnorm[i].resize( n_i );
255 _adj_m_unnorm[i].resize( n_i );
256 _new_adj_n[i].resize( n_i );
257 _new_adj_m[i].resize( n_i );
258 foreach( const Neighbor &I, _fg->nbV(i) ) {
259 { // calculate adj_n
260 Prob prod( _fg->factor(I).p() );
261 prod *= _adj_b_F_unnorm[I];
262 foreach( const Neighbor &j, _fg->nbF(I) )
263 if( i != j ) {
264 Prob n_jI( _bp_dual.msgN( j, j.dual ) );
265 const _ind_t &ind = _index( j, j.dual );
266 // multiply prod with n_jI
267 for( size_t x_I = 0; x_I < prod.size(); x_I++ )
268 prod[x_I] *= n_jI[ind[x_I]];
269 }
270 Prob marg( _fg->var(i).states(), 0.0 );
271 const _ind_t &ind = _index( i, I.iter );
272 for( size_t r = 0; r < prod.size(); r++ )
273 marg[ind[r]] += prod[r];
274 _new_adj_n[i][I.iter] = marg;
275 upMsgN( i, I.iter );
276 }
277
278 { // calculate adj_m
279 Prob prod( _adj_b_V_unnorm[i] );
280 assert( prod.size() == _fg->var(i).states() );
281 foreach( const Neighbor &J, _fg->nbV(i) )
282 if( J.node != I.node )
283 prod *= _bp_dual.msgM(i,J.iter);
284 _new_adj_m[i][I.iter] = prod;
285 upMsgM( i, I.iter );
286 }
287 }
288 }
289 }
290
291
292 void BBP::RegenerateSeqMessageAdjoints() {
293 size_t nv = _fg->nrVars();
294 _adj_m.resize( nv );
295 _adj_m_unnorm.resize( nv );
296 _new_adj_m.resize( nv );
297 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
298 size_t n_i = _fg->nbV(i).size();
299 _adj_m[i].resize( n_i );
300 _adj_m_unnorm[i].resize( n_i );
301 _new_adj_m[i].resize( n_i );
302 foreach( const Neighbor &I, _fg->nbV(i) ) {
303 // calculate adj_m
304 Prob prod( _adj_b_V_unnorm[i] );
305 assert( prod.size() == _fg->var(i).states() );
306 foreach( const Neighbor &J, _fg->nbV(i) )
307 if( J.node != I.node )
308 prod *= _bp_dual.msgM( i, J.iter );
309 _adj_m[i][I.iter] = prod;
310 calcUnnormMsgM( i, I.iter );
311 _new_adj_m[i][I.iter] = Prob( _fg->var(i).states(), 0.0 );
312 }
313 }
314 for( size_t i = 0; i < _fg->nrVars(); i++ ) {
315 foreach( const Neighbor &I, _fg->nbV(i) ) {
316 // calculate adj_n
317 Prob prod( _fg->factor(I).p() );
318 prod *= _adj_b_F_unnorm[I];
319 foreach( const Neighbor &j, _fg->nbF(I) )
320 if( i != j ) {
321 Prob n_jI( _bp_dual.msgN( j, j.dual) );
322 const _ind_t& ind = _index( j, j.dual );
323 // multiply prod with n_jI
324 for( size_t x_I = 0; x_I < prod.size(); x_I++ )
325 prod[x_I] *= n_jI[ind[x_I]];
326 }
327 Prob marg( _fg->var(i).states(), 0.0 );
328 const _ind_t &ind = _index( i, I.iter );
329 for( size_t r = 0; r < prod.size(); r++ )
330 marg[ind[r]] += prod[r];
331 sendSeqMsgN( i, I.iter,marg );
332 }
333 }
334 }
335
336
337 void BBP::calcNewN( size_t i, size_t _I ) {
338 _adj_psi_V[i] += T(i,_I) * _adj_n_unnorm[i][_I];
339 Prob &new_adj_n_iI = _new_adj_n[i][_I];
340 new_adj_n_iI = Prob( _fg->var(i).states(), 0.0 );
341 size_t I = _fg->nbV(i)[_I];
342 foreach( const Neighbor &j, _fg->nbF(I) )
343 if( j != i ) {
344 const Prob &p = _S[i][_I][j.iter];
345 const Prob &_adj_m_unnorm_jI = _adj_m_unnorm[j][j.dual];
346 LOOP_ij(
347 new_adj_n_iI[xi] += p[xij] * _adj_m_unnorm_jI[xj];
348 );
349 /* THE FOLLOWING WOULD BE ABOUT TWICE AS SLOW:
350 Var vi = _fg->var(i);
351 Var vj = _fg->var(j);
352 new_adj_n_iI = (Factor(VarSet(vi, vj), p) * Factor(vj,_adj_m_unnorm_jI)).marginal(vi,false).p();
353 */
354 }
355 }
356
357
358 void BBP::calcNewM( size_t i, size_t _I ) {
359 const Neighbor &I = _fg->nbV(i)[_I];
360 Prob p( U(I, I.dual) );
361 const Prob &adj = _adj_m_unnorm[i][_I];
362 const _ind_t &ind = _index(i,_I);
363 for( size_t x_I = 0; x_I < p.size(); x_I++ )
364 p[x_I] *= adj[ind[x_I]];
365 _adj_psi_F[I] += p;
366 /* THE FOLLOWING WOULD BE SLIGHTLY SLOWER:
367 _adj_psi_F[I] += (Factor( _fg->factor(I).vars(), U(I, I.dual) ) * Factor( _fg->var(i), _adj_m_unnorm[i][_I] )).p();
368 */
369
370 _new_adj_m[i][_I] = Prob( _fg->var(i).states(), 0.0 );
371 foreach( const Neighbor &J, _fg->nbV(i) )
372 if( J != I )
373 _new_adj_m[i][_I] += _R[I][I.dual][J.iter] * _adj_n_unnorm[i][J.iter];
374 }
375
376
377 void BBP::calcUnnormMsgN( size_t i, size_t _I ) {
378 _adj_n_unnorm[i][_I] = unnormAdjoint( _bp_dual.msgN(i,_I), _bp_dual.zN(i,_I), _adj_n[i][_I] );
379 }
380
381
382 void BBP::calcUnnormMsgM( size_t i, size_t _I ) {
383 _adj_m_unnorm[i][_I] = unnormAdjoint( _bp_dual.msgM(i,_I), _bp_dual.zM(i,_I), _adj_m[i][_I] );
384 }
385
386
387 void BBP::upMsgN( size_t i, size_t _I ) {
388 _adj_n[i][_I] = _new_adj_n[i][_I];
389 calcUnnormMsgN( i, _I );
390 }
391
392
393 void BBP::upMsgM( size_t i, size_t _I ) {
394 _adj_m[i][_I] = _new_adj_m[i][_I];
395 calcUnnormMsgM( i, _I );
396 }
397
398
399 void BBP::doParUpdate() {
400 for( size_t i = 0; i < _fg->nrVars(); i++ )
401 foreach( const Neighbor &I, _fg->nbV(i) ) {
402 calcNewM( i, I.iter );
403 calcNewN( i, I.iter );
404 }
405 for( size_t i = 0; i < _fg->nrVars(); i++ )
406 foreach( const Neighbor &I, _fg->nbV(i) ) {
407 upMsgM( i, I.iter );
408 upMsgN( i, I.iter );
409 }
410 }
411
412
413 void BBP::incrSeqMsgM( size_t i, size_t _I, const Prob &p ) {
414 if( props.clean_updates )
415 _new_adj_m[i][_I] += p;
416 else {
417 _adj_m[i][_I] += p;
418 calcUnnormMsgM(i, _I);
419 }
420 }
421
422
423 #if 0
424 Real pv_thresh=1000;
425 #endif
426
427
428 void BBP::updateSeqMsgM( size_t i, size_t _I ) {
429 if( props.clean_updates ) {
430 #if 0
431 if(_new_adj_m[i][_I].sumAbs() > pv_thresh ||
432 _adj_m[i][_I].sumAbs() > pv_thresh) {
433
434 DAI_DMSG("in updateSeqMsgM");
435 DAI_PV(i);
436 DAI_PV(_I);
437 DAI_PV(_adj_m[i][_I]);
438 DAI_PV(_new_adj_m[i][_I]);
439 }
440 #endif
441 _adj_m[i][_I] += _new_adj_m[i][_I];
442 calcUnnormMsgM( i, _I );
443 _new_adj_m[i][_I].fill( 0.0 );
444 }
445 }
446
447
448 void BBP::setSeqMsgM( size_t i, size_t _I, const Prob &p ) {
449 _adj_m[i][_I] = p;
450 calcUnnormMsgM( i, _I );
451 }
452
453
454 void BBP::sendSeqMsgN( size_t i, size_t _I, const Prob &f ) {
455 Prob f_unnorm = unnormAdjoint( _bp_dual.msgN(i,_I), _bp_dual.zN(i,_I), f );
456 const Neighbor &I = _fg->nbV(i)[_I];
457 assert( I.iter == _I );
458 _adj_psi_V[i] += f_unnorm * T( i, _I );
459 #if 0
460 if(f_unnorm.sumAbs() > pv_thresh) {
461 DAI_DMSG("in sendSeqMsgN");
462 DAI_PV(i);
463 DAI_PV(I);
464 DAI_PV(_I);
465 DAI_PV(_bp_dual.msgN(i,_I));
466 DAI_PV(_bp_dual.zN(i,_I));
467 DAI_PV(_bp_dual.msgM(i,_I));
468 DAI_PV(_bp_dual.zM(i,_I));
469 DAI_PV(_fg->factor(I).p());
470 }
471 #endif
472 foreach( const Neighbor &J, _fg->nbV(i) ) {
473 if( J.node != I.node ) {
474 #if 0
475 if(f_unnorm.sumAbs() > pv_thresh) {
476 DAI_DMSG("in sendSeqMsgN loop");
477 DAI_PV(J);
478 DAI_PV(f_unnorm);
479 DAI_PV(_R[J][J.dual][_I]);
480 DAI_PV(f_unnorm * _R[J][J.dual][_I]);
481 }
482 #endif
483 incrSeqMsgM( i, J.iter, f_unnorm * R(J, J.dual, _I) );
484 }
485 }
486 }
487
488
489 void BBP::sendSeqMsgM( size_t j, size_t _I ) {
490 const Neighbor &I = _fg->nbV(j)[_I];
491
492 // DAI_PV(j);
493 // DAI_PV(I);
494 // DAI_PV(_adj_m_unnorm_jI);
495 // DAI_PV(_adj_m[j][_I]);
496 // DAI_PV(_bp_dual.zM(j,_I));
497
498 size_t _j = I.dual;
499 const Prob &_adj_m_unnorm_jI = _adj_m_unnorm[j][_I];
500 Prob um( U(I, _j) );
501 const _ind_t &ind = _index(j, _I);
502 for( size_t x_I = 0; x_I < um.size(); x_I++ )
503 um[x_I] *= _adj_m_unnorm_jI[ind[x_I]];
504 um *= 1 - props.damping;
505 _adj_psi_F[I] += um;
506
507 /* THE FOLLOWING WOULD BE SLIGHTLY SLOWER:
508 _adj_psi_F[I] += (Factor( _fg->factor(I).vars(), U(I, _j) ) * Factor( _fg->var(j), _adj_m_unnorm[j][_I] )).p() * (1.0 - props.damping);
509 */
510
511 // DAI_DMSG("in sendSeqMsgM");
512 // DAI_PV(j);
513 // DAI_PV(I);
514 // DAI_PV(_I);
515 // DAI_PV(_fg->nbF(I).size());
516 foreach( const Neighbor &i, _fg->nbF(I) ) {
517 if( i.node != j ) {
518 const Prob &S = _S[i][i.dual][_j];
519 Prob msg( _fg->var(i).states(), 0.0 );
520 LOOP_ij(
521 msg[xi] += S[xij] * _adj_m_unnorm_jI[xj];
522 );
523 msg *= 1.0 - props.damping;
524 /* THE FOLLOWING WOULD BE ABOUT TWICE AS SLOW:
525 Var vi = _fg->var(i);
526 Var vj = _fg->var(j);
527 msg = (Factor(VarSet(vi,vj), S) * Factor(vj,_adj_m_unnorm_jI)).marginal(vi,false).p() * (1.0 - props.damping);
528 */
529 #if 0
530 if(msg.sumAbs() > pv_thresh) {
531 DAI_DMSG("in sendSeqMsgM loop");
532
533 DAI_PV(j);
534 DAI_PV(I);
535 DAI_PV(_I);
536 DAI_PV(_fg->nbF(I).size());
537 DAI_PV(_fg->factor(I).p());
538 DAI_PV(_S[i][i.dual][_j]);
539
540 DAI_PV(i);
541 DAI_PV(i.dual);
542 DAI_PV(msg);
543 DAI_PV(_fg->nbV(i).size());
544 }
545 #endif
546 assert( _fg->nbV(i)[i.dual].node == I );
547 sendSeqMsgN( i, i.dual, msg );
548 }
549 }
550 setSeqMsgM( j, _I, _adj_m[j][_I] * props.damping );
551 }
552
553
554 Real BBP::getUnMsgMag() {
555 Real s = 0.0;
556 size_t e = 0;
557 for( size_t i = 0; i < _fg->nrVars(); i++ )
558 foreach( const Neighbor &I, _fg->nbV(i) ) {
559 s += _adj_m_unnorm[i][I.iter].sumAbs();
560 s += _adj_n_unnorm[i][I.iter].sumAbs();
561 e++;
562 }
563 return s / e;
564 }
565
566
567 void BBP::getMsgMags( Real &s, Real &new_s ) {
568 s = 0.0;
569 new_s = 0.0;
570 size_t e = 0;
571 for( size_t i = 0; i < _fg->nrVars(); i++ )
572 foreach( const Neighbor &I, _fg->nbV(i) ) {
573 s += _adj_m[i][I.iter].sumAbs();
574 s += _adj_n[i][I.iter].sumAbs();
575 new_s += _new_adj_m[i][I.iter].sumAbs();
576 new_s += _new_adj_n[i][I.iter].sumAbs();
577 e++;
578 }
579 s /= e;
580 new_s /= e;
581 }
582
583 // tuple<size_t,size_t,Real> BBP::getArgMaxPsi1Adj() {
584 // size_t argmax_var=0;
585 // size_t argmax_var_state=0;
586 // Real max_var=0;
587 // for( size_t i = 0; i < _fg->nrVars(); i++ ) {
588 // pair<size_t,Real> argmax_state = adj_psi_V(i).argmax();
589 // if(i==0 || argmax_state.second>max_var) {
590 // argmax_var = i;
591 // max_var = argmax_state.second;
592 // argmax_var_state = argmax_state.first;
593 // }
594 // }
595 // assert(/*0 <= argmax_var_state &&*/
596 // argmax_var_state < _fg->var(argmax_var).states());
597 // return tuple<size_t,size_t,Real>(argmax_var,argmax_var_state,max_var);
598 // }
599
600
601 void BBP::getArgmaxMsgM( size_t &out_i, size_t &out__I, Real &mag ) {
602 bool found = false;
603 for( size_t i = 0; i < _fg->nrVars(); i++ )
604 foreach( const Neighbor &I, _fg->nbV(i) ) {
605 Real thisMag = _adj_m[i][I.iter].sumAbs();
606 if( !found || mag < thisMag ) {
607 found = true;
608 mag = thisMag;
609 out_i = i;
610 out__I = I.iter;
611 }
612 }
613 assert( found );
614 }
615
616
617 Real BBP::getMaxMsgM() {
618 size_t dummy;
619 Real mag;
620 getArgmaxMsgM( dummy, dummy, mag );
621 return mag;
622 }
623
624
625 Real BBP::getTotalMsgM() {
626 Real mag = 0.0;
627 for( size_t i = 0; i < _fg->nrVars(); i++ )
628 foreach( const Neighbor &I, _fg->nbV(i) )
629 mag += _adj_m[i][I.iter].sumAbs();
630 return mag;
631 }
632
633
634 Real BBP::getTotalNewMsgM() {
635 Real mag = 0.0;
636 for( size_t i = 0; i < _fg->nrVars(); i++ )
637 foreach( const Neighbor &I, _fg->nbV(i) )
638 mag += _new_adj_m[i][I.iter].sumAbs();
639 return mag;
640 }
641
642
643 Real BBP::getTotalMsgN() {
644 Real mag = 0.0;
645 for( size_t i = 0; i < _fg->nrVars(); i++ )
646 foreach( const Neighbor &I, _fg->nbV(i) )
647 mag += _adj_n[i][I.iter].sumAbs();
648 return mag;
649 }
650
651
652 void BBP::Regenerate() {
653 RegenerateInds();
654 RegenerateT();
655 RegenerateU();
656 RegenerateS();
657 RegenerateR();
658 RegenerateInputs();
659 RegeneratePsiAdjoints();
660 if( props.updates == Properties::UpdateType::PAR )
661 RegenerateParMessageAdjoints();
662 else
663 RegenerateSeqMessageAdjoints();
664 _iters = 0;
665 }
666
667
668 std::vector<Prob> BBP::getZeroAdjF( const FactorGraph &fg ) {
669 vector<Prob> adj_2;
670 adj_2.reserve( fg.nrFactors() );
671 for( size_t I = 0; I < fg.nrFactors(); I++ )
672 adj_2.push_back( Prob( fg.factor(I).states(), 0.0 ) );
673 return adj_2;
674 }
675
676
677 std::vector<Prob> BBP::getZeroAdjV( const FactorGraph &fg ) {
678 vector<Prob> adj_1;
679 adj_1.reserve( fg.nrVars() );
680 for( size_t i = 0; i < fg.nrVars(); i++ )
681 adj_1.push_back( Prob( fg.var(i).states(), 0.0 ) );
682 return adj_1;
683 }
684
685
686 void BBP::run() {
687 typedef BBP::Properties::UpdateType UT;
688 Real &tol = props.tol;
689 UT &updates = props.updates;
690
691 Real tic = toc();
692 switch( (size_t)updates ) {
693 case UT::SEQ_MAX: {
694 size_t i, _I;
695 Real mag;
696 do {
697 _iters++;
698 getArgmaxMsgM( i, _I, mag );
699 sendSeqMsgM( i, _I );
700 } while( mag > tol && _iters < props.maxiter );
701
702 if( _iters >= props.maxiter )
703 if( props.verbose >= 1 )
704 cerr << "Warning: BBP didn't converge in " << _iters << " iterations (greatest message magnitude = " << mag << ")" << endl;
705 break;
706 } case UT::SEQ_FIX: {
707 Real mag;
708 do {
709 _iters++;
710 mag = getTotalMsgM();
711 if( mag < tol )
712 break;
713
714 for( size_t i = 0; i < _fg->nrVars(); i++ )
715 foreach( const Neighbor &I, _fg->nbV(i) )
716 sendSeqMsgM( i, I.iter );
717 for( size_t i = 0; i < _fg->nrVars(); i++ )
718 foreach( const Neighbor &I, _fg->nbV(i) )
719 updateSeqMsgM( i, I.iter );
720 } while( mag > tol && _iters < props.maxiter );
721
722 if( _iters >= props.maxiter )
723 if( props.verbose >= 1 )
724 cerr << "Warning: BBP didn't converge in " << _iters << " iterations (greatest message magnitude = " << mag << ")" << endl;
725 break;
726 } case UT::SEQ_BP_REV:
727 case UT::SEQ_BP_FWD: {
728 const BP *bp = static_cast<const BP*>(_ia);
729 vector<pair<size_t, size_t> > sentMessages = bp->getSentMessages();
730 size_t totalMessages = sentMessages.size();
731 if( totalMessages == 0 ) {
732 cerr << "Asked for updates = " << updates << " but no BP messages; did you forget to set recordSentMessages?" << endl;
733 DAI_THROW(INTERNAL_ERROR);
734 }
735 if( updates==UT::SEQ_BP_FWD )
736 reverse( sentMessages.begin(), sentMessages.end() );
737 // DAI_PV(sentMessages.size());
738 // DAI_PV(_iters);
739 // DAI_PV(props.maxiter);
740 while( sentMessages.size() > 0 && _iters < props.maxiter ) {
741 // DAI_PV(sentMessages.size());
742 // DAI_PV(_iters);
743 _iters++;
744 pair<size_t, size_t> e = sentMessages.back();
745 sentMessages.pop_back();
746 size_t i = e.first, _I = e.second;
747 sendSeqMsgM( i, _I );
748 }
749 if( _iters >= props.maxiter )
750 if( props.verbose >= 1 )
751 cerr << "Warning: BBP updates limited to " << props.maxiter << " iterations, but using UpdateType " << updates << " with " << totalMessages << " messages" << endl;
752 break;
753 } case UT::PAR: {
754 do {
755 _iters++;
756 doParUpdate();
757 } while( (_iters < 2 || getUnMsgMag() > tol) && _iters < props.maxiter );
758 if( _iters == props.maxiter ) {
759 Real s, new_s;
760 getMsgMags( s, new_s );
761 if( props.verbose >= 1 )
762 cerr << "Warning: BBP didn't converge in " << _iters << " iterations (unnorm message magnitude = " << getUnMsgMag() << ", norm message mags = " << s << " -> " << new_s << ")" << endl;
763 }
764 break;
765 }
766 }
767 if( props.verbose >= 3 )
768 cerr << "BBP::run() took " << toc()-tic << " seconds " << doneIters() << " iterations" << endl;
769 }
770
771
772 double numericBBPTest( const InfAlg &bp, const vector<size_t> *state, const PropertySet &bbp_props, bbp_cfn_t cfn, double h ) {
773 // calculate the value of the unperturbed cost function
774 Real cf0 = getCostFn( bp, cfn, state );
775
776 // run BBP to estimate adjoints
777 BBP bbp( &bp, bbp_props );
778 initBBPCostFnAdj( bbp, bp, cfn, state );
779 bbp.run();
780
781 Real d = 0;
782 const FactorGraph& fg = bp.fg();
783
784 if( 1 ) {
785 // verify bbp.adj_psi_V
786
787 // for each variable i
788 for( size_t i = 0; i < fg.nrVars(); i++ ) {
789 vector<double> adj_est;
790 // for each value xi
791 for( size_t xi = 0; xi < fg.var(i).states(); xi++ ) {
792 // Clone 'bp' (which may be any InfAlg)
793 InfAlg *bp_prb = bp.clone();
794
795 // perturb it
796 size_t n = bp_prb->fg().var(i).states();
797 Prob psi_1_prb( n, 1.0 );
798 psi_1_prb[xi] += h;
799 // psi_1_prb.normalize();
800 size_t I = bp_prb->fg().nbV(i)[0]; // use first factor in list of neighbors of i
801 bp_prb->fg().factor(I) *= Factor( bp_prb->fg().var(i), psi_1_prb );
802
803 // call 'init' on the perturbed variables
804 bp_prb->init( bp_prb->fg().var(i) );
805
806 // run copy to convergence
807 bp_prb->run();
808
809 // calculate new value of cost function
810 Real cf_prb = getCostFn( *bp_prb, cfn, state );
811
812 // use to estimate adjoint for i
813 adj_est.push_back( (cf_prb - cf0) / h );
814
815 // free cloned InfAlg
816 delete bp_prb;
817 }
818 Prob p_adj_est( adj_est.begin(), adj_est.end() );
819 // compare this numerical estimate to the BBP estimate; sum the distances
820 cout << "i: " << i
821 << ", p_adj_est: " << p_adj_est
822 << ", bbp.adj_psi_V(i): " << bbp.adj_psi_V(i) << endl;
823 d += dist( p_adj_est, bbp.adj_psi_V(i), Prob::DISTL1 );
824 }
825 }
826 /* if(1) {
827 // verify bbp.adj_n and bbp.adj_m
828
829 // We actually want to check the responsiveness of objective
830 // function to changes in the final messages. But at the end of a
831 // BBP run, the message adjoints are for the initial messages (and
832 // they should be close to zero, see paper). So this resets the
833 // BBP adjoints to the refer to the desired final messages
834 bbp.RegenerateMessageAdjoints();
835
836 // for each variable i
837 for(size_t i=0; i<bp_dual.nrVars(); i++) {
838 // for each factor I ~ i
839 foreach(size_t I, bp_dual.nbV(i)) {
840 vector<double> adj_n_est;
841 // for each value xi
842 for(size_t xi=0; xi<bp_dual.var(i).states(); xi++) {
843 BP_dual bp_dual_prb(bp_dual);
844 // make h-sized change to newMsgN
845 bp_dual_prb.newMsgN(i,I)[xi] += h;
846 // recalculate beliefs
847 bp_dual_prb.CalcBeliefs();
848 // get cost function value
849 Real cf_prb = getCostFn(bp_dual_prb, cfn, &state);
850 // add it to list of adjoints
851 adj_n_est.push_back((cf_prb-cf0)/h);
852 }
853
854 vector<double> adj_m_est;
855 // for each value xi
856 for(size_t xi=0; xi<bp_dual.var(i).states(); xi++) {
857 BP_dual bp_dual_prb(bp_dual);
858 // make h-sized change to newMsgM
859 bp_dual_prb.newMsgM(I,i)[xi] += h;
860 // recalculate beliefs
861 bp_dual_prb.CalcBeliefs();
862 // get cost function value
863 Real cf_prb = getCostFn(bp_dual_prb, cfn, &state);
864 // add it to list of adjoints
865 adj_m_est.push_back((cf_prb-cf0)/h);
866 }
867
868 Prob p_adj_n_est(adj_n_est.begin(), adj_n_est.end());
869 // compare this numerical estimate to the BBP estimate; sum the distances
870 cerr << "i: " << i << ", I: " << I
871 << ", adj_n_est: " << p_adj_n_est
872 << ", bbp.adj_n(i,I): " << bbp.adj_n(i,I) << endl;
873 d += dist(p_adj_n_est, bbp.adj_n(i,I), Prob::DISTL1);
874
875 Prob p_adj_m_est(adj_m_est.begin(), adj_m_est.end());
876 // compare this numerical estimate to the BBP estimate; sum the distances
877 cerr << "i: " << i << ", I: " << I
878 << ", adj_m_est: " << p_adj_m_est
879 << ", bbp.adj_m(I,i): " << bbp.adj_m(I,i) << endl;
880 d += dist(p_adj_m_est, bbp.adj_m(I,i), Prob::DISTL1);
881 }
882 }
883 }
884 */
885 /* if(0) {
886 // verify bbp.adj_b_V
887 for(size_t i=0; i<bp_dual.nrVars(); i++) {
888 vector<double> adj_b_V_est;
889 // for each value xi
890 for(size_t xi=0; xi<bp_dual.var(i).states(); xi++) {
891 BP_dual bp_dual_prb(bp_dual);
892
893 // make h-sized change to b_1(i)[x_i]
894 bp_dual_prb._beliefs.b1[i][xi] += h;
895
896 // get cost function value
897 Real cf_prb = getCostFn(bp_dual_prb, cfn, &state);
898
899 // add it to list of adjoints
900 adj_b_V_est.push_back((cf_prb-cf0)/h);
901 }
902 Prob p_adj_b_V_est(adj_b_V_est.begin(), adj_b_V_est.end());
903 // compare this numerical estimate to the BBP estimate; sum the distances
904 cerr << "i: " << i
905 << ", adj_b_V_est: " << p_adj_b_V_est
906 << ", bbp.adj_b_V(i): " << bbp.adj_b_V(i) << endl;
907 d += dist(p_adj_b_V_est, bbp.adj_b_V(i), Prob::DISTL1);
908 }
909 }
910 */
911
912 // return total of distances
913 return d;
914 }
915
916
917 bool needGibbsState( bbp_cfn_t cfn ) {
918 switch( (size_t)cfn ) {
919 case bbp_cfn_t::CFN_GIBBS_B:
920 case bbp_cfn_t::CFN_GIBBS_B2:
921 case bbp_cfn_t::CFN_GIBBS_EXP:
922 case bbp_cfn_t::CFN_GIBBS_B_FACTOR:
923 case bbp_cfn_t::CFN_GIBBS_B2_FACTOR:
924 case bbp_cfn_t::CFN_GIBBS_EXP_FACTOR:
925 return true;
926 default:
927 return false;
928 }
929 }
930
931
932 void initBBPCostFnAdj( BBP &bbp, const InfAlg &ia, bbp_cfn_t cfn_type, const vector<size_t> *stateP ) {
933 const FactorGraph &fg = ia.fg();
934
935 switch( (size_t)cfn_type ) {
936 case bbp_cfn_t::CFN_BETHE_ENT: {
937 vector<Prob> b1_adj;
938 vector<Prob> b2_adj;
939 vector<Prob> psi1_adj;
940 vector<Prob> psi2_adj;
941 b1_adj.reserve( fg.nrVars() );
942 psi1_adj.reserve( fg.nrVars() );
943 b2_adj.reserve( fg.nrFactors() );
944 psi2_adj.reserve( fg.nrFactors() );
945 for( size_t i = 0; i < fg.nrVars(); i++ ) {
946 size_t dim = fg.var(i).states();
947 int c = fg.nbV(i).size();
948 Prob p(dim,0.0);
949 for( size_t xi = 0; xi < dim; xi++ )
950 p[xi] = (1 - c) * (1 + log( ia.beliefV(i)[xi] ));
951 b1_adj.push_back( p );
952
953 for( size_t xi = 0; xi < dim; xi++ )
954 p[xi] = -ia.beliefV(i)[xi];
955 psi1_adj.push_back( p );
956 }
957 for( size_t I = 0; I < fg.nrFactors(); I++ ) {
958 size_t dim = fg.factor(I).states();
959 Prob p( dim, 0.0 );
960 for( size_t xI = 0; xI < dim; xI++ )
961 p[xI] = 1 + log( ia.beliefF(I)[xI] / fg.factor(I).p()[xI] );
962 b2_adj.push_back( p );
963
964 for( size_t xI = 0; xI < dim; xI++ )
965 p[xI] = -ia.beliefF(I)[xI] / fg.factor(I).p()[xI];
966 psi2_adj.push_back( p );
967 }
968 bbp.init( b1_adj, b2_adj, psi1_adj, psi2_adj );
969 break;
970 } case bbp_cfn_t::CFN_FACTOR_ENT: {
971 vector<Prob> b2_adj;
972 b2_adj.reserve( fg.nrFactors() );
973 for( size_t I = 0; I < fg.nrFactors(); I++ ) {
974 size_t dim = fg.factor(I).states();
975 Prob p( dim, 0.0 );
976 for( size_t xI = 0; xI < dim; xI++ ) {
977 double bIxI = ia.beliefF(I)[xI];
978 if( bIxI < 1.0e-15 )
979 p[xI] = -1.0e10;
980 else
981 p[xI] = 1 + log( bIxI );
982 }
983 b2_adj.push_back(p);
984 }
985 bbp.init( bbp.getZeroAdjV(fg), b2_adj );
986 break;
987 } case bbp_cfn_t::CFN_VAR_ENT: {
988 vector<Prob> b1_adj;
989 b1_adj.reserve( fg.nrVars() );
990 for( size_t i = 0; i < fg.nrVars(); i++ ) {
991 size_t dim = fg.var(i).states();
992 Prob p( dim, 0.0 );
993 for( size_t xi = 0; xi < fg.var(i).states(); xi++ ) {
994 double bixi = ia.beliefV(i)[xi];
995 if( bixi < 1.0e-15 )
996 p[xi] = -1.0e10;
997 else
998 p[xi] = 1 + log( bixi );
999 }
1000 b1_adj.push_back( p );
1001 }
1002 bbp.init( b1_adj );
1003 break;
1004 } case bbp_cfn_t::CFN_GIBBS_B:
1005 case bbp_cfn_t::CFN_GIBBS_B2:
1006 case bbp_cfn_t::CFN_GIBBS_EXP: {
1007 // cost functions that use Gibbs sample, summing over variable marginals
1008 vector<size_t> state;
1009 if( stateP == NULL )
1010 state = getGibbsState( ia, 2*ia.Iterations() );
1011 else
1012 state = *stateP;
1013 assert( state.size() == fg.nrVars() );
1014
1015 vector<Prob> b1_adj;
1016 b1_adj.reserve(fg.nrVars());
1017 for( size_t i = 0; i < state.size(); i++ ) {
1018 size_t n = fg.var(i).states();
1019 Prob delta( n, 0.0 );
1020 assert(/*0<=state[i] &&*/ state[i] < n);
1021 double b = ia.beliefV(i)[state[i]];
1022 switch( (size_t)cfn_type ) {
1023 case bbp_cfn_t::CFN_GIBBS_B:
1024 delta[state[i]] = 1.0;
1025 break;
1026 case bbp_cfn_t::CFN_GIBBS_B2:
1027 delta[state[i]] = b;
1028 break;
1029 case bbp_cfn_t::CFN_GIBBS_EXP:
1030 delta[state[i]] = exp(b);
1031 break;
1032 default:
1033 DAI_THROW(INTERNAL_ERROR);
1034 }
1035 b1_adj.push_back( delta );
1036 }
1037 bbp.init( b1_adj );
1038 break;
1039 } case bbp_cfn_t::CFN_GIBBS_B_FACTOR:
1040 case bbp_cfn_t::CFN_GIBBS_B2_FACTOR:
1041 case bbp_cfn_t::CFN_GIBBS_EXP_FACTOR: {
1042 // cost functions that use Gibbs sample, summing over factor marginals
1043 vector<size_t> state;
1044 if( stateP == NULL )
1045 state = getGibbsState( ia, 2*ia.Iterations() );
1046 else
1047 state = *stateP;
1048 assert( state.size() == fg.nrVars() );
1049
1050 vector<Prob> b2_adj;
1051 b2_adj.reserve( fg.nrVars() );
1052 for( size_t I = 0; I < fg.nrFactors(); I++ ) {
1053 size_t n = fg.factor(I).states();
1054 Prob delta( n, 0.0 );
1055
1056 size_t x_I = getFactorEntryForState( fg, I, state );
1057 assert(/*0<=x_I &&*/ x_I < n);
1058
1059 double b = ia.beliefF(I)[x_I];
1060 switch( (size_t)cfn_type ) {
1061 case bbp_cfn_t::CFN_GIBBS_B_FACTOR:
1062 delta[x_I] = 1.0;
1063 break;
1064 case bbp_cfn_t::CFN_GIBBS_B2_FACTOR:
1065 delta[x_I] = b;
1066 break;
1067 case bbp_cfn_t::CFN_GIBBS_EXP_FACTOR:
1068 delta[x_I] = exp( b );
1069 break;
1070 default:
1071 DAI_THROW(INTERNAL_ERROR);
1072 }
1073 b2_adj.push_back( delta );
1074 }
1075 bbp.init( bbp.getZeroAdjV(fg), b2_adj );
1076 break;
1077 } default:
1078 DAI_THROW(UNKNOWN_ENUM_VALUE);
1079 }
1080 }
1081
1082
1083 Real getCostFn( const InfAlg &ia, bbp_cfn_t cfn_type, const vector<size_t> *stateP ) {
1084 double cf = 0.0;
1085 const FactorGraph &fg = ia.fg();
1086
1087 switch( (size_t)cfn_type ) {
1088 case bbp_cfn_t::CFN_BETHE_ENT: // ignores state
1089 cf = -ia.logZ();
1090 break;
1091 case bbp_cfn_t::CFN_VAR_ENT: // ignores state
1092 for( size_t i = 0; i < fg.nrVars(); i++ )
1093 cf += -ia.beliefV(i).entropy();
1094 break;
1095 case bbp_cfn_t::CFN_FACTOR_ENT: // ignores state
1096 for( size_t I = 0; I < fg.nrFactors(); I++ )
1097 cf += -ia.beliefF(I).entropy();
1098 break;
1099 case bbp_cfn_t::CFN_GIBBS_B:
1100 case bbp_cfn_t::CFN_GIBBS_B2:
1101 case bbp_cfn_t::CFN_GIBBS_EXP: {
1102 assert( stateP != NULL );
1103 vector<size_t> state = *stateP;
1104 assert( state.size() == fg.nrVars() );
1105 for( size_t i = 0; i < fg.nrVars(); i++ ) {
1106 double b = ia.beliefV(i)[state[i]];
1107 switch( (size_t)cfn_type ) {
1108 case bbp_cfn_t::CFN_GIBBS_B:
1109 cf += b;
1110 break;
1111 case bbp_cfn_t::CFN_GIBBS_B2:
1112 cf += b * b / 2.0;
1113 break;
1114 case bbp_cfn_t::CFN_GIBBS_EXP:
1115 cf += exp( b );
1116 break;
1117 default:
1118 DAI_THROW(INTERNAL_ERROR);
1119 }
1120 }
1121 break;
1122 } case bbp_cfn_t::CFN_GIBBS_B_FACTOR:
1123 case bbp_cfn_t::CFN_GIBBS_B2_FACTOR:
1124 case bbp_cfn_t::CFN_GIBBS_EXP_FACTOR: {
1125 assert( stateP != NULL );
1126 vector<size_t> state = *stateP;
1127 assert( state.size() == fg.nrVars() );
1128 for( size_t I = 0; I < fg.nrFactors(); I++ ) {
1129 size_t x_I = getFactorEntryForState( fg, I, state );
1130 double b = ia.beliefF(I)[x_I];
1131 switch( (size_t)cfn_type ) {
1132 case bbp_cfn_t::CFN_GIBBS_B_FACTOR:
1133 cf += b;
1134 break;
1135 case bbp_cfn_t::CFN_GIBBS_B2_FACTOR:
1136 cf += b * b / 2.0;
1137 break;
1138 case bbp_cfn_t::CFN_GIBBS_EXP_FACTOR:
1139 cf += exp( b );
1140 break;
1141 default:
1142 DAI_THROW(INTERNAL_ERROR);
1143 }
1144 }
1145 break;
1146 } default:
1147 cerr << "Unknown cost function " << cfn_type << endl;
1148 DAI_THROW(UNKNOWN_ENUM_VALUE);
1149 }
1150 return cf;
1151 }
1152
1153
1154 } // end of namespace dai
1155
1156
1157 /* {{{ GENERATED CODE: DO NOT EDIT. Created by
1158 ./scripts/regenerate-properties include/dai/bbp.h src/bbp.cpp
1159 */
1160 namespace dai {
1161
1162 void BBP::Properties::set(const PropertySet &opts)
1163 {
1164 const std::set<PropertyKey> &keys = opts.allKeys();
1165 std::set<PropertyKey>::const_iterator i;
1166 bool die=false;
1167 for(i=keys.begin(); i!=keys.end(); i++) {
1168 if(*i == "verbose") continue;
1169 if(*i == "maxiter") continue;
1170 if(*i == "tol") continue;
1171 if(*i == "damping") continue;
1172 if(*i == "updates") continue;
1173 if(*i == "clean_updates") continue;
1174 cerr << "BBP: Unknown property " << *i << endl;
1175 die=true;
1176 }
1177 if(die) {
1178 DAI_THROW(UNKNOWN_PROPERTY_TYPE);
1179 }
1180 if(!opts.hasKey("verbose")) {
1181 cerr << "BBP: Missing property \"verbose\" for method \"BBP\"" << endl;
1182 die=true;
1183 }
1184 if(!opts.hasKey("maxiter")) {
1185 cerr << "BBP: Missing property \"maxiter\" for method \"BBP\"" << endl;
1186 die=true;
1187 }
1188 if(!opts.hasKey("tol")) {
1189 cerr << "BBP: Missing property \"tol\" for method \"BBP\"" << endl;
1190 die=true;
1191 }
1192 if(!opts.hasKey("damping")) {
1193 cerr << "BBP: Missing property \"damping\" for method \"BBP\"" << endl;
1194 die=true;
1195 }
1196 if(!opts.hasKey("updates")) {
1197 cerr << "BBP: Missing property \"updates\" for method \"BBP\"" << endl;
1198 die=true;
1199 }
1200 if(!opts.hasKey("clean_updates")) {
1201 cerr << "BBP: Missing property \"clean_updates\" for method \"BBP\"" << endl;
1202 die=true;
1203 }
1204 if(die) {
1205 DAI_THROW(NOT_ALL_PROPERTIES_SPECIFIED);
1206 }
1207 verbose = opts.getStringAs<size_t>("verbose");
1208 maxiter = opts.getStringAs<size_t>("maxiter");
1209 tol = opts.getStringAs<double>("tol");
1210 damping = opts.getStringAs<double>("damping");
1211 updates = opts.getStringAs<UpdateType>("updates");
1212 clean_updates = opts.getStringAs<bool>("clean_updates");
1213 }
1214 PropertySet BBP::Properties::get() const {
1215 PropertySet opts;
1216 opts.Set("verbose", verbose);
1217 opts.Set("maxiter", maxiter);
1218 opts.Set("tol", tol);
1219 opts.Set("damping", damping);
1220 opts.Set("updates", updates);
1221 opts.Set("clean_updates", clean_updates);
1222 return opts;
1223 }
1224 string BBP::Properties::toString() const {
1225 stringstream s(stringstream::out);
1226 s << "[";
1227 s << "verbose=" << verbose << ",";
1228 s << "maxiter=" << maxiter << ",";
1229 s << "tol=" << tol << ",";
1230 s << "damping=" << damping << ",";
1231 s << "updates=" << updates << ",";
1232 s << "clean_updates=" << clean_updates;
1233 s << "]";
1234 return s.str();
1235 }
1236 } // end of namespace dai
1237 /* }}} END OF GENERATED CODE */
libDAI - A free/open source C++ library for Discrete Approximate Inference methods